Latest / Strachey Lectures / Strachey Lecture: Use or Be Used - Regaining Control of AI

Transcript

Strachey Lecture: Use or Be Used - Regaining Control of AI · Strachey Lectures

• 0:12
  I'd like to welcome you to the Trinity Church, Straight G lecture,
• 0:15
  the Straight G lectures O the distinguished lecture series from the Department of Computer Science.
• 0:21
  We have a hash tag which is in front of me. Hash ox, straight Tilak.
• 0:26
  Please do tweet at me. Nice about it. If you are if you are going to tweet.
• 0:30
  Before we begin, I would like to acknowledge our sponsors.
• 0:33
  Oxford Asset Management now for a number of years have been sponsoring these lectures and it's been a transformative thing for us,
• 0:41
  enabling us to bring in speakers and to put on a show that we simply wouldn't have been able to do otherwise.
• 0:48
  So we're extremely grateful to Oxford Asset Management for their continuing support.
• 0:53
  I I'm sure they would like me to point out that they are hiring.
• 0:57
  So if you Google Oxford Asset Management, you will hire, you will find many job opportunities.
• 1:02
  But thank you to Oxford Asset Management. So now on to our speakers and I'm delighted to welcome Neil Laurence.
• 1:10
  Neil is the inaugural DeepMind Professor of Machine Learning at the University of Cambridge, where he leads the university's flagship mission on A.I.
• 1:18
  A.I. At camp. He's been working on machine learning models for 20 years,
• 1:23
  and he recently returned to academia after three years of director of Machine Learning Amazon.
• 1:28
  His main interest is the interaction of machine learning with the physical world,
• 1:32
  and this interest was triggered by deploying machine learning solutions in the African context, where end to end solutions are essential.
• 1:40
  And this has inspired new research directions at the interface of machine learning and systems research.
• 1:45
  And that work is funded by Senior A.I. Fellowship from the Alan Turing Institute.
• 1:50
  He's interim chair of the Advisory Board of the UK Centre for Data Ethics and Innovation and a member of the UK's A.I. Council,
• 1:58
  which advises government on all matters. A.I. Neil is a visiting professor at the University of Sheffield,
• 2:05
  where he was for a number of years as an academic and is the co-host of Talking Machines.
• 2:10
  And the title of Neil's lecture could hardly be more contemporary use or be used.
• 2:17
  Regaining Control of Artificial Intelligence. Males, The floor is yours.
• 2:29
  Thanks to this chip that's working. Okay. And thank you for that wonderful introduction.
• 2:33
  And thank you, Mike, Leslie, everyone, for the invitation and an apology.
• 2:39
  This invitation was originally for a year ago, and I caught COVID just before I was due to give the presentation.
• 2:48
  The title is actually the same, so imagine how much better the world would have been had I been able to give this lecture a year ago.
• 2:57
  Look, another another consequence of the pandemic.
• 3:01
  Okay. So I try and use historic.
• 3:05
  I'm not going to give a technical thought really today. I'm just going to talk about issues that lead to technical questions.
• 3:11
  And when I think about artificial intelligence, I like to sort of use these three apocryphal quotes that will be in the talk today.
• 3:19
  So the apocryphal quote about Henry Ford is when?
• 3:23
  And then he says that if I'd asked people what they wanted out of a car, they would have asked for faster horse.
• 3:30
  And that provokes in my mind the question. So if you ask people what do they want out of artificial intelligence, is the answer a smarter human?
• 3:39
  And then the question that follows is, is even what does that mean?
• 3:43
  I think, by the way, if we had a faster horse, we'd already have autonomous vehicles.
• 3:47
  So that would have been a pretty cool thing. So maybe the customer's always right.
• 3:51
  Now, I talk about this a lot and I'm only going to go through it very briefly because I just want to contextualise the way I think about it,
• 4:00
  which is when we look at humans in human intelligence, I've been trying to do public understanding, machine learning folks for about ten years,
• 4:08
  and I wanted to capture why machine intelligence is so very different from human intelligence and
• 4:14
  simply thought at one level it's because I'm speaking to you using sound and machines communicate,
• 4:21
  using radio waves or light speed and cables.
• 4:24
  So they communicate about a million times faster.
• 4:26
  And this leads to all these sort of consequences that for me to talk to you if I'm talking in a good rate of speech,
• 4:34
  if we look at how much information I'm sharing, it's about 2000 bits per minute.
• 4:39
  So that's a bit as defined by Shannon, an information theory that's the equivalent information as 2000 coin tosses.
• 4:46
  So it's about telling you the result of 2000 tennis matches where the odds were even every minute, which is pretty good now.
• 4:55
  But to do that, I'm sharing with you the thoughts are inside my head, and it's very hard to estimate how many calculations per seconds we're doing.
• 5:02
  But if you look at the best estimates I could find for simulating a brain,
• 5:06
  you would need a machine about as fast as the Met Office supercomputer down in Exeter,
• 5:12
  which can do about 1,000,000,000 billion calculations per second.
• 5:17
  Now, if you compare that to a simple computer, these slides actually date back a few years.
• 5:21
  So maybe they're a bit faster than this. Nowadays, a computer can communicate at 60 billion bits per minute if it's using gigabit per second Internet.
• 5:30
  So that means if we turn out, that's a nine to salary.
• 5:34
  So imagine your university student stipend is $20 a month.
• 5:39
  It's not great money, but it's kind of maybe enough to survive on.
• 5:43
  And then imagine your university stipend is $60 billion a month.
• 5:49
  It's actually closer to the budget of a country, so it's closer to what the UK spends on the whole of health and Social Security $60 billion a month.
• 5:59
  So it's operating in terms of this information exchange, much, much faster.
• 6:02
  But then our best estimate is that a typical machine in terms of number of calculations, it can do per second.
• 6:09
  And I've done discriminant calculations a second distance. You see there are about 100 billion calculations.
• 6:15
  So actually a typical machine is doing a lot less compute per second than our brains.
• 6:20
  Of course, we can't access all that compute. Neither can the computer access all the compute it does.
• 6:25
  So imagine if I wanted to share with you 1 seconds worth of calculations from my head.
• 6:31
  If these numbers are right, it would take me a billion years to tell you all the calculations.
• 6:37
  5 billion years to tell you all the calculations that had occurred in my head in 1 seconds.
• 6:42
  So while my battery is running low, this is an. I'm not sure I was referring to my brain or the machine.
• 6:50
  Not me. I can see a blob that's not strong. So I meant to infer human.
• 6:59
  Human intelligence. So, 5 billion years.
• 7:04
  What is a computer? What did you share? 1 seconds worth of computation.
• 7:07
  You would take it about 20 minutes. So this means your intelligence is locked in.
• 7:11
  So what I mean by locked in is you can hear these stories.
• 7:14
  And sometimes in a longer version of this section, I would talk about a guy called Jean-Dominique Bobic,
• 7:19
  who was the French editor of Elle magazine, who was locked in, could only communicate by winking.
• 7:24
  They made a film about him. He wrote the book about the use the film after he was locked in.
• 7:29
  And I think that's a state we all think about. We all think, oh, wow, what would it be like to be in that state?
• 7:33
  Well, that's the state you're in, right? You just evolve that way.
• 7:37
  You don't notice. But compared to the machine, that's where you're at.
• 7:42
  So in some sense, it's already a nonsense to talk about a more intelligent human being,
• 7:48
  because the human being is defined by constraints, not by capabilities.
• 7:53
  We're defined in terms of our intelligence, in the way we share information by a limited ability to communicate.
• 7:58
  And we've evolved in that state for hundreds of thousands of years,
• 8:02
  as Homo sapiens should say that in the Natural History Museum and for, you know, millions of years as primates, etc., etc.
• 8:09
  And all animals do this to some. Well, mammals, higher mammals do this to someone great, to all that sort of thing.
• 8:14
  So second apocryphal quote, the six word novel never written by Ernest Hemingway,
• 8:18
  but apocryphal written by honest Hemingway for sale baby shoes never wore.
• 8:24
  Now, so that six words, 7 to 2 bits of information on average, 12 bits per word we have.
• 8:30
  But there's so much more meaning in those 12 words than 72 bits of information.
• 8:38
  And that's because you exist as humans. You sit within a wider culture, a wider society.
• 8:43
  And you can imagine what it would feel to be the human that has to place that up.
• 8:50
  So we use this information bandwidth in a way that allows us to lean on the fact that we are all human and we understand what it means to be a human,
• 9:01
  to communicate things that feel like there's a lot more than 72 bits of information in there,
• 9:05
  because we can all imagine what it felt like to write that advert.
• 9:10
  So although the information but I'm sure our context is very important and we
• 9:14
  can extract a lot of information from very few provided pieces of information.
• 9:20
  So does anyone know who this apocryphal quote is? Apocryphal divine.
• 9:24
  There are three types of three families and statistics. It is Mark Twain.
• 9:32
  Not apocryphal in Mark Twain, but Mark Twain appropriately credits it to Benjamin Disraeli.
• 9:37
  But it was said by Lord Balfour, the foreign secretary tried to do it to someone else at some point.
• 9:42
  But the point about this quote is that one of the reasons that we end up misunderstanding and misinterpreting
• 9:48
  statistics is because we're so information constrained that when we get little bits of information,
• 9:54
  we tend to overinterpret. We tend to see more behind that data than we have any right to.
• 10:01
  And that comes out in a bad way in statistics.
• 10:03
  It comes out in a good way when you're imagining what it felt for the person that had to write that advert.
• 10:09
  So I kind of apocryphal.
• 10:11
  Benjamin Disraeli I tend to think one of the the modern equivalent of these quotes is this There are three types of lies, lie families and big data.
• 10:21
  Because what's happened in this modern data era and that's being driven by the machine interconnection,
• 10:27
  is requiring more and more information, more and more data.
• 10:31
  And we're basically forgetting the hard lessons of a field known as mathematical statistics.
• 10:37
  So that quote from Disraeli predates the mathematical field of statistics.
• 10:41
  So I'm quite close to computer science, government statistics.
• 10:44
  They study mathematical statistics. They just don't say mathematical statistics.
• 10:48
  They study that the they say the process of understanding whether whether you can believe what the statistics is purporting to say or not.
• 10:57
  And we don't even say mathematical statistics anymore.
• 10:59
  But it was a new field that had to come in to stop people misinterpreting statistics, which is the sort of science mistake.
• 11:06
  So that's Karl Pearson, of course, unfortunately associated with that field is a lot of fairly unpleasant things like eugenics.
• 11:13
  Most of the early mathematical statisticians were also involved in eugenics to one form or another, because they believe once you could tell this,
• 11:21
  we should be looking to improve the human species in some way, and that led to some quite problematic results.
• 11:27
  So that is still an open question about what does mathematical data science look like in our era.
• 11:33
  And it's being triggered by this sort of new image because we now have an evolved relationship with information.
• 11:41
  So in the past, very, very far past 100,000 years ago, we could go in the museum and check,
• 11:49
  but I think we would been in groups of about 30 to 50 operating together,
• 11:53
  hunter gatherers, maybe slightly larger, trusting each other, basing sort of inferences about what's going on.
• 12:00
  Those relationships of trust and understanding who each of us is. Since then, we became we came into cities.
• 12:07
  We've now grown things. We have to introduce laws to govern us.
• 12:11
  And we also ended up with statistics to look at how those cities are being managed.
• 12:16
  And we've had to teach ourselves not to misinterpret those statistics when making decisions.
• 12:20
  We removed ourselves from that direct contact of person to person, and that's led to this evolved relationship with information.
• 12:28
  So in the past 100 years, you know, this is the situation that you had humans with this very low bandwidth connection.
• 12:37
  And a lot of what we were trying to do was manage that low bandwidth connection
• 12:40
  to ensure that we weren't misinterpreting the information we were receiving.
• 12:45
  But now what we've got is this We've got an enormously high bandwidth connection between
• 12:50
  the machine and data and then a low bandwidth connection between us and the machine.
• 12:56
  And this is leading to all sorts of problems. You know, and I we can we can look back over the last ten years in terms of election manipulation,
• 13:05
  in terms of social media, etc., etc., and what that's done.
• 13:09
  And you can put that down to the fact that the machine is now mediating what information it chooses to show.
• 13:14
  I mean, it's actually using classical statistics to work out how to draw us in to our engagement with the machine.
• 13:22
  But it's not using classical physics to say, how can I share the truth with this individual?
• 13:27
  It's using classical statistics to say, how can I cause this individual to engage with me?
• 13:31
  And of course, the answer turns out to be share information with them that, you know, fulfils their prejudices, that type of thing.
• 13:40
  She's not great.
• 13:42
  So what we have is this as the challenge for data science, that we're trying to deal with this new high bandwidth connection to the machine,
• 13:50
  which is then presenting information to us in the low bandwidth way, and we've got classical statistics being the traditional way now.
• 13:57
  The first thing I want to emphasise is when you see this connection, you realise, well,
• 14:02
  the statistics have a massive role in mediating these new connections and understanding what's going wrong.
• 14:06
  But it's not everything. But there is enormous role of computer science as well.
• 14:11
  You know, mathematics just in understanding this, this new flow of information.
• 14:17
  But then the thing I think we need to also be talking about on top of that is this new wave of movements which are actually,
• 14:25
  I would argue, operating almost directly on that link between the human and the machine.
• 14:29
  So forming new ways for us to communicate with the machine.
• 14:34
  So in the past, what a statistician would do if they wanted to represent data is they would train the human to
• 14:40
  understand what a P value is or apparently misunderstand what a P value is or what an error bar is,
• 14:47
  or how to interpret a graph.
• 14:49
  And then the machine would share, you know, if you were using the machine that information in the form of a graph and provide that information to you.
• 14:57
  What we've got now is machines that will just talk to us and when they talk to us,
• 15:03
  they are going right to the depths of that desire to overinterpret because that feeding into our desire to anthropomorphise eyes.
• 15:13
  And I can't say anthropomorphise very well. So I just say anthrax.
• 15:18
  So we all tend to treat complex things as if they are human in some form, whether it's our cat or whether it's our car or whether it's our computer.
• 15:28
  And when machines are starting to exhibit language, that tendency is going to increase even more.
• 15:35
  But the problem is that we mustn't forget this fundamental difference, that these intelligences are quite different from us.
• 15:40
  They are quite alien to ours. They can augment us.
• 15:45
  But the point in this tool is they can never replace us and they should never be allowed to replace us,
• 15:52
  because at the core of our intelligence is our limitations.
• 15:56
  And those are things that right from its birth, however capable the machine has, it will never have those limitations baked into it.
• 16:05
  Okay, so when I think about the scale of what we're experiencing now,
• 16:10
  I was giving I was convening the UK, AI Fellows, students and postdocs at a hackathon,
• 16:17
  partially because I strongly believe that we are so dependent on that generation
• 16:22
  of people to deploy and use these technologies in the best way possible,
• 16:28
  because it's not going to be the old people like myself who understand the ways in which these things can be deployed and used.
• 16:34
  We can see that in every previous revolution it's young people. So I've tried to come up with analogies for how big I think this revolution is.
• 16:41
  And sometimes I talk all right about the printing press and the sort of extraordinary effect that had 500 years ago leading to the Enlightenment,
• 16:49
  but also sort of 500 years of war around religion and all sorts of things going on in Europe.
• 16:54
  But then I kind of caution to the printing press doesn't go far enough.
• 16:57
  So the revolution I think we're engaged in at the moment is arguably akin to the revolution of rights.
• 17:04
  So this is a sort of early cuneiform tablet that if you go back to well, I started working with a colleague,
• 17:12
  geologist in Cambridge who works on Nepal from 1500 B.C. to 1000 B.C. and reads these tablets with,
• 17:20
  you know, court cases written down this time that as far as I can make out,
• 17:23
  the Greeks were some form of barbarians going around the Mediterranean, attacking cities.
• 17:29
  And at the same time, the city of Nepal has 2000 years of written history going back to, you know, before Gilgamesh existed.
• 17:38
  So this is a society and a culture that existed some 3000 years ago that had 2000 years of written history,
• 17:47
  written in the form of these sort of clay tablets.
• 17:49
  But then the thing I sort of think about is what was it like for the first person when when the Epic of Gilgamesh, which was an oral tradition,
• 17:56
  like presumably most of these traditions were like The Odyssey and everything else was suddenly written down for the first time,
• 18:02
  and the sort of bard who's learned this whole epic and can sing it and do everything else.
• 18:06
  He turns up and says, Now this or I may I put it written down, Now don't need you anymore.
• 18:11
  We don't need you to remember things because we can remember things.
• 18:16
  All a piece of tabloid. And the first form of usage for these tablets was not to write down poems.
• 18:22
  It was to write down accounts. So-and-so owes me three sheep.
• 18:26
  I owe them four lots of call. And I think there's a tremendous parallel between that and the computer,
• 18:32
  which so far has mainly been and somehow a tool of the accountants excel spreadsheets.
• 18:37
  It doesn't and can't interface with us in the poetic language we might like to use and what it suddenly gains.
• 18:45
  I mean it. Pearson But it's not something.
• 18:47
  This has been going on for a number of years, but it's crossed the threshold into, Oh yes, you can talk to the machine.
• 18:55
  And I think that this is really one particular reason why this is sending society into these sort of spasms of
• 19:03
  excitement is is related to an experience I had in 2017 where we were doing a Royal Society report on machine learning.
• 19:12
  So people were thinking about this a long time ago and worrying about how technology would change things.
• 19:17
  And we were very focussed at the time as it was on my Richard and Daniel Susskind on the future profession.
• 19:22
  It's a very different technological transform, the work of human experts,
• 19:25
  and we were doing this work from 2015 to about 2017 and about midway through the work when we were coming up with
• 19:31
  a lot of conclusions about how the lives of the middle classes were going to be affected by this revolution,
• 19:40
  we had a particular occurrence that really stuck with me and that was the Brexit vote and we were in the wrong society.
• 19:48
  About a week later and we were sat around a table and we were all digesting the consequences of this vote.
• 19:55
  And there were employees from all parties and there were policy experts and there were domain experts.
• 20:01
  And one of the MP said, Well, I bet no one in this room voted for Brexit.
• 20:05
  And it suddenly occurred to me, I'll bet your rights.
• 20:09
  And isn't that the problem?
• 20:11
  Because the people who voted for Brexit are the ones who were disengaged from the decision making, and that's why they voted for Brexit.
• 20:20
  And you can see that in populism wherever it exists. And that gave me a great fear because this book speaks to the empowerment.
• 20:28
  It speaks to all of us that are already engaged in decision making.
• 20:33
  And that's also what you see going on now. Now, I'm not saying the allies aren't going to be transformed as educated people,
• 20:41
  but my experience is that educated people are amongst the best equipped to handle such change.
• 20:49
  And mean what I have in my head is that of the coin pushing machine.
• 20:54
  So this is a coin which I used to love before I as a kid, I don't know if anyone knows what they are.
• 20:59
  These are the more fair grounds. You've got to bring the Blackpool or whatever else in you.
• 21:03
  You've got these two pennies and they go on the sets and the sets move in and out and you draw the
• 21:07
  coin and the coin sort of goes in the sand here and then it lands and it should push these coins off.
• 21:12
  You can see these coins are deformed and then when the step pushes out, some coins will fall off and hopefully you'll win.
• 21:18
  Now, which coins fall off? Is it the coins that move? No, it's the coins on the edge.
• 21:23
  Because those people who are already on the edge of society, who are already challenged by things,
• 21:30
  are the ones that are going to be affected by this form of disruption.
• 21:33
  And my sense is, yes, those of us in the professions, those of us who are educated can do things well.
• 21:39
  We'll have to learn. We'll be disrupted in a major way. But we actually have the intelligence and capability to do your jobs.
• 21:45
  It's the knock on effects that it has on those people that have already given up
• 21:49
  that I met a guy in Lancaster once who used to who decided not to go to university,
• 21:55
  wanted to be in a band. He went catering, called into catering. He worked at the prison.
• 21:59
  The president got shot.
• 22:00
  And then, of course, I met him because he was driving a taxi, because that's where, if you ask any taxi driver why they're driving a taxi,
• 22:06
  you'll always get this sequence of stories about, you know, because anyone can sort of drive a taxi.
• 22:11
  But there are lots of people's where there isn't much demand for taxi.
• 22:16
  So when we did have a call for the Royal Society back in 2017, it was called machine made the complex computers that lead by example.
• 22:23
  I still think it's a really valid report that is actually still referenced today,
• 22:27
  but one of the things we did is we spoke to people and we asked them what they would like machine learning solutions to do.
• 22:34
  And so we did a it was the Ipsos MORI poll on public views, machine learning.
• 22:38
  So this is qualitative research, which I think is going to be increasingly important in this domain.
• 22:43
  And I understand in qualitative research, even if you're a quantitative scientist, needs to improve dramatically.
• 22:48
  So these are the sort of things they said, and I'm not sure if you can read, but these are sort of areas we looked at.
• 22:53
  So we asked about health and health and they can see the greatest potential benefit to individuals and society.
• 23:00
  We asked about social care,
• 23:03
  so they saw the potential for resourcing issues that they fear to mobilise machines with due to involvement and emotional contact.
• 23:10
  Right. So that's clearly a worry and sort of thing. But they could see that there was benefit, the best case scenario, unachievable.
• 23:18
  She and they were keeping carers to spend more time with patients.
• 23:22
  I mean that would be brilliant if that's what we've done in any of our hospitals or social care.
• 23:26
  I don't think it's a very good record.
• 23:28
  But marketing maybe even thought marketing was, Oh yeah, we could be used to tailoring marketing and that might not be a bad idea.
• 23:34
  Transport driver's cars could have benefits.
• 23:37
  These are just regular people.
• 23:39
  They were sort of in Birmingham in a sort of like small groups finance universe and thought about our to monitor potentially fraudulent activity.
• 23:46
  People are really, really smart about what they want technology to do.
• 23:50
  Participants tend to think machine learning and crime the thought patterns in.
• 23:54
  Good idea. So let's see how might work accurately in practice.
• 23:57
  Smart people, they saw it as a useful tool to aid the military police resources but concerned about consequence to stereotyping.
• 24:03
  Individual people are smart. This is why at a time, by the way, when most people had not even heard what machine learning was,
• 24:09
  that was the first thing you had to do is educate them what machine learning was.
• 24:13
  Public education partisans are concerned that tainted education based on machine learning
• 24:17
  would result in killing and limiting people to some career path that too young and again,
• 24:21
  really, really smart. But they also sort of see benefits in that.
• 24:25
  What is the one area that all participants thought was ridiculous?
• 24:30
  That we should not be trying to do that? They could see no point in it whatsoever.
• 24:35
  RS They still fail to see the purpose of machine learning and read the poetry for all.
• 24:41
  Other case study.
• 24:42
  Participants recognise that machine might not do a better job than a human, however, but they do not think this application creates art,
• 24:48
  as doing so is considered to be a fundamental human activity that machines could only mimic best.
• 24:54
  Where have we made the greatest progress in machine learning in the ensuing six years?
• 25:01
  In mimicking what humans do.
• 25:03
  So in honour of that, I often get to go right to Silicon Valley That introduces a large language model us spoken by Mickey McHugh.
• 25:11
  This was chapter 54. Chapter 3.5 didn't do quite as good a job, but yes,
• 25:19
  I don't know who pockets to enable new POC is as if woven by is POC inappropriate person to create?
• 25:24
  Largely because of different play, but it's not a fixed.
• 25:28
  Right. Well, is the crossover plan. It's the one where they create large language models and there's roaming now.
• 25:36
  And you know, it's in nowadays. You know, you've got better call Saul, all sorts going on.
• 25:41
  The characters appear in other things. Digital Athenaeum.
• 25:46
  My goodness. But what would that fundamentally be?
• 25:51
  So I want to say, does anyone recognise what this map is of?
• 25:56
  You're not. What's. What's the map of. Western Europe.
• 26:01
  Yeah. And can you guess what year it is and what the circumstances of the matter.
• 26:06
  Well, go to that site and can you guess the date? There's only a few significant dates in World War Two before, but again, in 44 June, Portugal.
• 26:17
  Yeah, June the fifth, 1944. And can you guys see my penis?
• 26:23
  As a former teacher, it's a job and not of the weather stations.
• 26:28
  Yeah. So it's the pressure and the different motivations. And and actually, you know, I get it from the ice baths.
• 26:35
  So what was going on in June 44 was that there was a storm in the channel.
• 26:39
  What was also going on was that the tides were at the right level for an invasion of Europe to occur.
• 26:44
  What was difficult for the Germans is that the weather comes from the West and that way the stations were on France.
• 26:53
  Does anyone know where Rommel was on the day of the invasion at his wife's 50th birthday party in Germany, because he got the weather report and said,
• 27:02
  Well, there's no way there'll be an invasion tomorrow so I can go home to my wife's birthday party and everything will be fine.
• 27:10
  Of course, he was kind of rudely awoken when I say and had to drive rapidly back to the front.
• 27:16
  Now, if we look on the alternative map, this is the allied view of the weather on that day.
• 27:22
  And the allies actually on the same day, the front says, and they understand that there's going to be a gap in the weather.
• 27:26
  And that's actually multiple weather reports that are in conflict in this even controversy
• 27:30
  today over who agreed that there was going to be a significant gap in the weather.
• 27:34
  But the interesting thing about weather forecasting in both these cases is it was being done by interpolation.
• 27:40
  So you had a number of weather stations spaced across Britain.
• 27:44
  I mean, actually it's interesting to me. Martin I'm not sure saying we should have had weather information from France because of Enigma decrypts,
• 27:53
  but whether they would put it on the map or not.
• 27:54
  So we could actually interpolate across the channel when we did our weather reports, whereas the Germans,
• 27:59
  I don't think, could have some weather information on the previous map, I think in Cornwall.
• 28:03
  So I'm not sure how they're getting it and what the CORNISH doing.
• 28:11
  But basically you're having to interpret. But for Britain, you're interpreting in the right direction because the weather is moving in the right way.
• 28:19
  A modern weather forecast is, of course, done using Navier-stokes equation on that big computer done in Exeter,
• 28:25
  and they use Navier-stokes equation that allows them to extrapolate. And that extrapolation is sort of key in giving us these long term predictions
• 28:32
  up to five or six days because we understand the physics of the weather.
• 28:35
  We're not just looking at what the pressure is instead of having a stab at where it might move through some minimal sort of examination over time.
• 28:43
  So that's why the Germans got the weather wrong. Now, what's that got to do with TATP?
• 28:48
  Well, fundamentally, all it's doing is it's operating. It doesn't have access.
• 28:53
  It's in this position. It doesn't have access to the future where humans are given this tool and they
• 28:59
  start to work with this tool and they start doing what they will do with that.
• 29:04
  Why does that come about? Well, the enormous quantity of data we fed these machines is basically at the primary level.
• 29:12
  It's giving a feature space in which interpolation works.
• 29:16
  And all these capabilities like logical reasoning stuff are emergent because that
• 29:21
  feature space is emergent by just looking at enormous quantities of our data.
• 29:25
  But it like in the final layer of this thing, it's still having to interpolate.
• 29:30
  It can't build that feature space and extrapolate in such a way that says how well, how would McCue show actually react?
• 29:38
  You know, what would they do, first of all, which you could see it could say something,
• 29:42
  but we'll only know what McCue would do when the modern McCue shows get given the tool.
• 29:46
  And so at this very key point, given that humans, whatever their limitations, are coming out, this problem from particular perspective,
• 29:55
  you can never replace them because as soon as you augment them with the tool,
• 29:59
  they are something that is far greater than either themselves or the tool on its own.
• 30:03
  Just like we became far greater than ourselves by writing things in clay tablets, you know, for a large amount of time.
• 30:10
  That was how we shared knowledge and built upon knowledge and understood geometry and all sorts of other things.
• 30:15
  So those models are only interpolated now.
• 30:20
  This is something I talk about a lot, and I should not question what I would have said about this a little bit.
• 30:26
  Because when people talk about artificial intelligence, a question occurred to me is what do they mean?
• 30:32
  Because different people mean different things.
• 30:35
  I was at the Natural History Museum in London talking with some of the scientists there the other week and on my way,
• 30:41
  and I got chatting to the security guard. And I always like to ask people, what do you think it is?
• 30:47
  And it's quite hard to understand. A common thing is behind this because intelligence is such an emotive word.
• 30:54
  So they sort of think it's like them. It's like, well, it can't be quite like you.
• 30:57
  But I do think want to say to you that this technology might and could do is be the first technology that when it automates,
• 31:06
  tries to adapt to who we are or I think that's what it purports to do.
• 31:10
  Right. So all previous generations of automation, if you invented a weaving in a room or if you invented a spinning journey,
• 31:17
  it required everyone to turn up at the factory and service the machine.
• 31:21
  So Samuel Butler. It's about this in the 19th century,
• 31:24
  the way in which these forms of automation actually enslave us to the machine because we're adaptable and the machine isn't.
• 31:31
  So if you look at railways, you know, you all have to turn up on time station.
• 31:34
  You didn't even have universal time across the United Kingdom until they had railways.
• 31:40
  So every previous generation of automation requires us as flexible entities to adapt to the machine.
• 31:46
  And I think what people expect that they're getting from an AI is an entity that will adapt to them.
• 31:54
  I think it's a fallacy, but maybe I'm wrong.
• 31:56
  And maybe what we're seeing now is when you have machines that can understand the whole of human history and culture,
• 32:02
  they can mimic adaptation to a certain it will feel a bit more.
• 32:05
  I don't know if we'll truly feel that in two, three years time. I think people are feeling that today.
• 32:10
  I can have a conversation with a machine about my software that feels like it is adapting to who I am and what I like.
• 32:18
  But when it comes to then applying that in critical decision making,
• 32:23
  so a lot of what we care about in society is what one might think of as consequential decisions being made by people about us and who we are.
• 32:31
  I think there's kind of a problem with this point of view because the thing I sort of started noticing,
• 32:36
  and I think it's also written about I'm not the first to notice it and people use different words,
• 32:40
  but everyone in machine learning talks about we want decisions that are fair.
• 32:45
  But when they say that, we want fair decisions. I've noticed that people mean a couple of different things.
• 32:53
  And in fact,
• 32:53
  it was Karl Rasmussen that he went out for a run once because I was talking about the complexity of a fair decision and how difficult it is.
• 33:02
  If I'm going to make a nuanced decision about some individuals in front of me from
• 33:07
  different backgrounds who all appear qualified to enter the University of Cambridge,
• 33:11
  and I'm taking into account all their different backgrounds, what they've done up until now, etc., etc., because they can all manage the degree.
• 33:19
  I'm using an immense amount of my understanding of these individuals as a human being and trying to find out about who they are.
• 33:26
  And it seems fair that I should do that because I'm taking into account these individuals
• 33:30
  experience of life and the struggles they've had to get through to get there.
• 33:34
  But it's another form of fairness that says, no, the rules should be clear. Now.
• 33:38
  Basically, you should have to get these grades and then you should be accepted into university.
• 33:43
  And the same thing applies for any consequential decision making, like giving loans or, you know, court cases.
• 33:51
  And this is now reflected in the General Data Protection regulation,
• 33:55
  which says that there are certain protected characteristics and you should not ever decide about someone on the basis of those characteristics.
• 34:03
  But what I find interesting about this is it's always like this is political alley.
• 34:07
  We decide which side, which politic politics is, errs on which side.
• 34:12
  To me, both of these extremes are dystopias. The extreme where all you're doing is considering everyone's history and and if only they be given
• 34:20
  the chances and giving everyone the same opportunity is one dystopia and the extreme way you say,
• 34:26
  no, this is the fixed rule, and only if you pass this rule do you get to have the beneficial outcome or the negative outcome is another dystopia.
• 34:36
  And in practice, we're always operating somewhere between.
• 34:40
  And it was sort of struggling to think about this because it's really important.
• 34:44
  How can you have something that is procedurally fair and simultaneously nuanced and substantively fair?
• 34:51
  Because it's something to be perceived.
• 34:53
  To be fair, we all have to understand, like, I don't think the tax code is fair because no one understands the tax code.
• 34:59
  So if you pay someone a large amount of money, they can find a hole in the tax code and you can pay less tax.
• 35:05
  So that's not procedurally fair, even though it's a process. Right.
• 35:08
  And it's not procedure fair because it's difficult to understand. So for something to be perceived fair, it has to be clear to everyone.
• 35:14
  But simultaneously, we're going to take into account all these nuances.
• 35:18
  How can we have something that is taking into account nuances and also be simple?
• 35:23
  Well, the answer goes back to that word.
• 35:25
  And folks, if you gather a group of people together, if you convene them really well and teach them how to make decisions,
• 35:32
  you can have a decision making process that is simultaneously nuanced and procedurally fair,
• 35:39
  because you can have people I mean, the most complex things in this room are the other people in this room.
• 35:46
  But we all think we have a deep and intuitive understanding of those people.
• 35:49
  I mean, when we do action because we co-evolved together, right? So that goes right to the heart of who we are as humans.
• 35:55
  So I call this a marvellous resolution between these two extremes that the only way you
• 36:00
  can break these two is to have a procedure where you've got well trained human beings.
• 36:05
  And of course, very often we're not well trained with bias or whatever.
• 36:08
  If you train us humans well about good decision making and then you also have a procedure that sort of says, yeah,
• 36:14
  we can how to get these marks, and then you'll be before a committee, and then that committee will make the final decision.
• 36:19
  Everyone says, Sure, that sounds fine if you say deal terms of.
• 36:22
  Say. And then the machine looks at your entire history and does really complicated calculations
• 36:27
  about what's going on that that's very hard for people to understand that sort of procedure,
• 36:32
  because it's a procedure, but one that we don't feel an intuition about unless we're assuming that machine is operating like you.
• 36:39
  So for me, this says that for consequential decision making,
• 36:42
  in order to have decisions that can balance between these two forms of fairness, we must always have humans in the loop.
• 36:51
  But what does that mean? Because we know machines can be a great help in these decisions.
• 36:58
  Okay.
• 36:58
  So I kind of think that the other side to that and I love these lectures, I listen to them recently, these are the best lectures on air quality here.
• 37:06
  And from 2002 and then on Al, because what I, Norah O'Neill is talking about is a question of trust.
• 37:13
  And she's talking about a world where everything has become more procedural and we're calling everything to account through processes.
• 37:21
  And she's talking about the way that erodes our professional faith in individuals who are paid to make good decisions.
• 37:29
  Now, there are clearly things that we all know about professional individuals who have made corrupted decisions that are in the papers all the time.
• 37:35
  But what Neil is saying is that by increasing accountability in terms of scoring them and marking them and making measures,
• 37:42
  you're actually eroding that professional duty.
• 37:45
  And what she talks about a lot is that all rights come with duties and that you can't have rights with duties.
• 37:52
  And what you will increasingly here, certainly if you hear this, is lapses again,
• 37:57
  is is that that's kind of the way that people talk about things like certainly even in academic politics,
• 38:03
  I very often don't hear my colleagues talking about what their duties are towards their students,
• 38:08
  but I often hear them talking about what their rights are around their intellectual property.
• 38:12
  And it would be good if they were focusing a bit more on what their duties are.
• 38:16
  And the point O'Neill is making is you can't have rights without duties now.
• 38:22
  I mean, she can't have duties because the machine itself does not participate in the society.
• 38:27
  It can't be trusted in the same way. And that's the point I Nora O'Neill makes.
• 38:31
  And it's a very strong point that sticks with me,
• 38:33
  because the first time she made that point was when I was interviewing her for the Royal Society Report,
• 38:38
  and I asked her if we could trust the machine. And she told me very specifically, you couldn't.
• 38:43
  But listening to her ideas of why that is the case is really about this is the same thing.
• 38:48
  You need humans with a human understanding of what professional duty is to be in the loop around this decision making.
• 38:55
  So what does that look like?
• 38:56
  So they actually she found a quote that relates to this example that I was using before I even commissioned for the elections.
• 39:03
  So again, universities are, I think I could say, on the basis of ability and promise,
• 39:07
  but they're supposed to also administration will represent the intent. There's no guarantee that the process meets the target.
• 39:13
  And I mean, she doesn't say this.
• 39:14
  What I'm trying to say is that the only way you can do that is if you have trusted individuals who are balancing those things as part of that process.
• 39:21
  And the only way you can trust those individuals is they're well trained and well convened,
• 39:25
  etc., etc., and they're held to account by their position in society.
• 39:28
  You can't trust the machine to do that because it doesn't have a position in society.
• 39:33
  So what does this mean for in terms of what we're doing with artificial intelligence today?
• 39:38
  Well, I kind of think as a result, artificial intelligence could only ever be seen as a tool.
• 39:44
  And if anyone is telling you it's a replacement for us, I think that's a ridiculous notion.
• 39:48
  That always has to be a ridiculous notion. Apart from the following, what does it mean for it to be a tool?
• 39:55
  I mean, I that's I want to talk about is the National Advisory Committee of Aeronautics in Langley Field,
• 40:02
  who participated in the Second World War and before that in testing aircraft and when they would test the aircraft.
• 40:09
  One of the things they did, a guy called Bob Gilruth wrote, is the handbook on how an aircraft should handle.
• 40:16
  So what he did and he's a member, the test pilots,
• 40:18
  they had what was called spectrum came out of I don't know which one, but I read a lot about seven come out race 47.
• 40:25
  When they got a new plan they went up and flew these plans and they moved the
• 40:29
  stick around and the plane did things and pilots would say flies like a blue,
• 40:33
  fly like a brick, whatever. But they didn't know what that meant in terms of how they should change the plane to fly that, sir.
• 40:41
  So what Bob Gilbert did is he wrote a paper that you can go and read it online somewhere
• 40:45
  that talks about how a plane should respond when you put £10 of pressure on the stick.
• 40:51
  Right. So if you put £10 of pressure on the left and it's a fighter, it should roll at a certain rate.
• 40:55
  If you put £10 of stick pressure on the stick and it's a bomber, it should roll at a slower rate As you approach your stall,
• 41:01
  the section JUDDER on the run into the stall to warn the pilot.
• 41:05
  He quantified all of this. Right.
• 41:07
  So when these pilots were flying planes, they could feel the plane and the plane was communicating in a way that was well understood.
• 41:16
  Now that control of a stick. Well, the other thing about nationalise the Commission Aeronautics, it became a founding group of method.
• 41:24
  And the people that did this work also designed the crewed capsules, the Mercury capsule, the Gemini capsule, and the Apollo capsule.
• 41:32
  The actually indulged in providing that feedback to pilots and the control to pilots on entities that were entirely controlled by computers.
• 41:40
  So the Apollo guidance computer was the thing that was firing the thrusters.
• 41:44
  But Neil Armstrong is moving the stick. So in that case, you've got this really interesting situation where the.
• 41:49
  Neil Armstrong is not in control because the computer's in control, but Neil Armstrong feels like he's in control.
• 41:57
  And that's the kind of balance we're looking for, because for the things that Neil Armstrong cares about, he is in control.
• 42:04
  And the same is true for the pilot of an A380 today.
• 42:07
  When they move that stick,
• 42:09
  the plane that's going into a computer and the computer is deciding what to do and sometimes the computer overrules the pilot.
• 42:15
  It's also true for flight fighter pilots today where fighter pilots would not be capable of flying, obviously, without the computer systems.
• 42:22
  So the notion of control you're looking for is something that's akin to this,
• 42:26
  where you want the intent of the human being to be understood and projected and enhanced by the machine.
• 42:32
  But the machine itself is doing things that are beyond the capability of any given human.
• 42:37
  Now, this is incredibly difficult,
• 42:38
  and I find this is a really ironic that the control stick was first developed by the Wright Brothers in 1904 or whenever they flew for their flyer,
• 42:47
  and it just went forward and backwards because they could only go up and down.
• 42:50
  And within 36 years we characterise that interface.
• 42:54
  The interface we're now faced with now is one, two.
• 42:57
  Let's just talk to our machines and our ability to characterise that interface between the computer and the human is minimal,
• 43:09
  like in terms of how we can be manipulated or how we should present uncertain information or how
• 43:16
  we share an understanding of a patient with a machine or how it should share it back is minimal.
• 43:23
  We don't have the equivalent of £10 left should lead to this and we urgently need to
• 43:28
  get on top of that because so far we've been manipulated in fairly simplistic ways.
• 43:33
  Like here's a like button or here's an image you'll like or click on this, or here's a question about We'll never believe what so-and-so looked like.
• 43:40
  Now, when you knew them in the 1980s, apparently you will take on that.
• 43:45
  That's simplistic.
• 43:46
  But now there's all sorts of opportunities for machines to interact with this in this highly complex way that we haven't quantified.
• 43:53
  So I kind of think that the main thing that we need is the equivalent of a proving ground where when we're looking at these models,
• 44:00
  we can understand how they are being deployed in practice.
• 44:05
  So understand the nature of the tool, what's the potential, what it pitfalls.
• 44:09
  And that's the way we can build cyber capability.
• 44:12
  There are going to be loads of disruptions, but there's loads of things we need to do and loads of ways we can be empowered.
• 44:18
  So the question is how do we get this understanding as widespread as possible?
• 44:23
  Like I said in the introduction, one of the things that has been enormously inspiration to me is working on point data science solutions in Africa.
• 44:31
  So Data Science Africa is a bottom up initiative for capacity building.
• 44:35
  They've signed machinery and they are on the African continent. And what I love about it is you're always trying to deploy end to end solutions.
• 44:42
  So you don't sit in an office and imagine what a farmer might want or a clinician might want.
• 44:47
  You actually go out to the health centre or the field and you don't imagine
• 44:51
  what the Ministry of Health might want or Minister of Agriculture might want.
• 44:54
  You actually go and talk to the Ministry of Agriculture and you find most things that we think about machine learning are utterly irrelevant and the
• 45:01
  most of the challenges that people are genuinely experiencing are not being properly researched because they're all at the interfaces with people.
• 45:09
  So the thing we need to urgently do is to get on top of working with those people.
• 45:13
  So there's an awful mess. In 2015, when we first set up weeks of talks about how we got in and finally get some key facts on the alarm.
• 45:23
  So you'll find all these notes and everything. Another aspect to this is how do people maintain power?
• 45:29
  How do people maintain control of the system in a societal way?
• 45:33
  Well, actually, all this information is coming from us as individuals.
• 45:37
  It's our collective eyes data that is being put into these machines.
• 45:40
  And we have rights over that data,
• 45:42
  whether they are intellectual property rights or whether they're coming from the general Data Protection regulation.
• 45:46
  But one challenge is that there's an asymmetry of power.
• 45:49
  So if I want to call up Google and say, Can you stop doing this with my data, they're not going to answer my call.
• 45:53
  So the notion of a data trust and there's a whole family of mainstream media is now is an
• 45:57
  entity which collectivise these people's data rights and operates on this ecosystem to say,
• 46:03
  well, basically you can't do that with these models because we want you to delete our data if you're going to do that.
• 46:09
  So there's a lot of work that's been going on. We've got some in this area reports 18 months of their trust initiative, 200 experts,
• 46:17
  some 70 or so and exciting free real world data transfer funding as pilot Wonga,
• 46:23
  which is a sort of local data trust people around bricks in the definition two of which are associated with medical data,
• 46:30
  one with people who opted out for the general practitioners recent data sharing scheme
• 46:36
  and another one for the Born in Scotland cohort saying which is just starting out.
• 46:41
  The other thing I think is vitally important and this is our funding,
• 46:45
  we also have to make funding here from Schmidt futures for a project money for four years now is the accelerated programme for scientific discovery,
• 46:53
  where the main focus is to go out to data. Science to go out to scientific domain experts and talk to them about their scientific problems
• 47:02
  and try and work out what the solutions they need from a machine learning perspective are.
• 47:06
  Because only by understanding the problems they face is it's entirely inspired by the
• 47:11
  talk to the farmer in the field approach that the DSA has to take the South Africa.
• 47:17
  But working with scientists and we go beyond science to engineering and humanities got project,
• 47:22
  as I said in my serology friend to try and understand what is it they're looking at?
• 47:26
  How how would they like their lives to be better?
• 47:29
  Rather than building a neural network and imposing on people, talking to people about the war, the problems you are facing.
• 47:35
  And it's exactly inspired by the fact that nurses aren't getting to spend more time with patients through computers.
• 47:40
  They're getting to spend more time with computers. And the reason is because that computer solution is imposed from the centre,
• 47:46
  rather than by talking to the nurse about what they need and these new capabilities in terms of our ability to code and communicate with machines,
• 47:55
  offer an opportunity to turn that around. And this stuff about there won't be any jobs, there's a lot of jobs doing that.
• 48:02
  We are utterly failed to deliver on this for like the last eight years, as we've just seen from what people ask for from the Royal Society poll.
• 48:09
  Of course, jobs will be different and of course it will be disruptive.
• 48:13
  So the final thing is that we're trying to sort of bring that together with the CAM flagship mission.
• 48:20
  And if you read and do feel free, we published this report before we got the mission approved,
• 48:24
  because the thing I hate about grant proposals is when they've rejected it and you don't get to do anything with your proposal.
• 48:30
  So I said, Well, well, we're going to write a report first, publish it, and then you can decide whether to fund it or not.
• 48:36
  Part of the reason for that is because, rightly or wrongly, people look to institutions like Cambridge and Oxford to provide leadership.
• 48:43
  So this report brought together a number of these strands of thinking communication across the university.
• 48:48
  What people expect to see from our and, you know, we hope other people will copy it,
• 48:52
  because what it's trying to do is start from what the societal challenges are at MIT,
• 48:57
  that Cambridge is not positioned to solve all societal challenges,
• 49:01
  but it's position to be a partner in helping people solve those challenges and try and get people to think about how we contribute in that way.
• 49:08
  Because it's up to leading institutions like Oxford and Cambridge and Imperial and UCL and
• 49:12
  Sheffield and Manchester to really stop all those coins falling off the side of the step.
• 49:19
  So Professor following that is just convening and I love it.
• 49:23
  To see what's so cool about this is relatively small amounts of funding.
• 49:28
  If you're funding the right people, it tends to be young people in an interdisciplinary way.
• 49:32
  You can just see beautiful things happen. I think we had this whole meeting where we just funded these groups to the tune of £10,000,
• 49:38
  and I was thinking, Yeah, that's about the cost of a business class flight.
• 49:42
  And yet some of these groups had produced like workshops, working papers, discussions that were just thinking in ways that only young people can.
• 49:51
  The old people who've been stuck in this area for too long are just not seeing the possibilities.
• 49:56
  And so it's really vital. We saw funding at that level.
• 49:59
  And with that, I'll just say thanks. I don't have any conclusions other than as the enormous amount of work to do, but we're going to do it together.
• 50:08
  And if we do it together, well, I think we can lead to a society that is much better for everyone,
• 50:14
  not just the professionals, but also all those people on the edge.
• 50:18
  Thank you very much.